Stochastic Equilibrium Problems arising in the energy industry

Transcription

1 Stochastic Equilibrium Problems arising in the energy industry Claudia Sagastizábal (visiting researcher IMPA) ENEC workshop, IPAM, Los Angeles, January 2016 joint work with J.P. Luna (UFRJ) and M. Solodov (IMPA)

2 What this talk is about? For equilibrium models of energy markets (including stochastic versions with risk aversion); + Modelling issues + Existence and other theoretical issues + Solution techniques

3 Energy markets can be large Strategic sectors: subject to regulations in quality, price and entry couple several regions and markets Electric Power: (source EPEX)

4 Energy markets can be large Strategic sectors: subject to regulations in quality, price and entry couple several regions and markets Natural Gas: Energy Policy, 36: , Egging, Gabriel, Holtz, Zhuang, A complementarity model for the European natural gas market

5 Market: Premises + Agents (producers, traders, logistics) -take unilateral decisions -behave competitively + A representative of the consumers (the ISO) -focuses on the benefits of consumption -seeking a price that matches supply and demand -while keeping prices low + Agents actions coupled by some relations, clearing the market.

6 Market: Premises + Agents (producers, traders, logistics) -take unilateral decisions -behave competitively + A representative of the consumers (the ISO) -focuses on the benefits of consumption -seeking a price that matches supply and demand -while keeping prices low + Agents actions coupled by some relations, clearing the market, (MC)

7 Market: Premises + Agents (producers, traders, logistics) -take unilateral decisions -behave competitively + A representative of the consumers (the ISO) -focuses on the benefits of consumption -seeking a price that matches supply and demand -while keeping prices low + Agents actions coupled by some relations, clearing the market, (MC) Today, models from game theory or complementarity leading to Variational Inequalities (VIs) (i.e., sufficiently "convex")

8 Do Different models yield different decisions? Mixed Complementarity formulations Agents maximize profit independently Supply Demand: Market Clearing constraint (MC) multiplier equilibrium price Price is an exogenous concave function of the total offer: π = π( i qi ) Models from game theory Agents maximize profit independently Supply Demand: Market Clearing constraint (MC) multiplier equilibrium price

9 Do Different models yield different decisions? Mixed Complementarity formulations Agents maximize profit independently Supply Demand: Market Clearing constraint (MC) multiplier equilibrium price Price is an exogenous concave function of the total offer: π = π( i qi ) Models from game theory Agents maximize profit independently market clearing constraint (supply=demand)

10 Do Different models yield different decisions? Mixed Complementarity formulations Agents maximize profit independently Supply Demand: Market Clearing constraint (MC) multiplier equilibrium price Price is an exogenous concave function of the total offer: π = π( i qi ) Models from game theory Agents minimize cost s.t. MC MC multiplier (variational) equilibrium price

11 Do Different models yield different decisions? Mixed Complementarity formulations Agents maximize profit independently Supply Demand: Market Clearing constraint (MC) multiplier equilibrium price Price is an exogenous concave function of the total offer: π = π( i qi ) Models from game theory Agents minimize cost s.t. MC MC multiplier (variational) equilibrium price

12 Do Different models yield different decisions? Mixed Complementarity formulations Agents maximize profit independently Supply Demand: Market Clearing constraint (MC) multiplier equilibrium price Price is an exogenous concave function of the total offer: π = π( i qi ) Models from game theory Agents minimize cost s.t. MC How different are these models? MC multiplier (variational) equilibrium price Consumers indirectly represented Notation: q = (q i,q i ), in particular π = π(q i,q i )

13 Market: Example as a Mixed Complementarity Problem + Agents (producers, traders, logistics) ith producer problem max r i (q i ) s.t. q i Q i + Revenue r i (q i ) = π q i c i (q i ) (price is an exogenous function of all the offer) + Agents actions coupled by a market clearing constraint + Equilibrium price coincides with the exogenous

14 Market: Example as a Mixed Complementarity Problem + Agents (producers, traders, logistics) ith producer problem max r i (q i,q i ) minc i (q i,q i )(c i = r i ) s.t. q i Q i + Revenue r i (q i ) = π q i c i (q i ) = r i (q i,q i ) (price is an exogenous function π(q) of all the offer) + Agents actions coupled by a market clearing constraint + Equilibrium price coincides with the exogenous

15 Market: Example as a Mixed Complementarity Problem + Agents (producers, traders, logistics) ith producer problem max r i (q i,q i ) minc i (q i,q i )(c i = r i ) s.t. q i Q i + Revenue r i (q i ) = π q i c i (q i ) = r i (q i,q i ) (price is an exogenous function π(q) of all the offer) + Agents actions coupled by a market clearing constraint MC(q i,q i ) = 0 (mult. π) + Equilibrium price coincides with the exogenous

16 Market: Example as a Mixed Complementarity Problem + Agents (producers, traders, logistics) ith producer problem max r i (q i,q i ) minc i (q i,q i )(c i = r i ) s.t. q i Q i + Revenue r i (q i ) = π q i c i (q i ) = r i (q i,q i ) (price is an exogenous function π(q) of all the offer) + Agents actions coupled by a market clearing constraint MC(q i,q i ) = 0 (mult. π) + Equilibrium price π coincides with the exogenous π( q)

17 Market: Equilibrium price: π Mixed Complementarity Model Agents problems min c i (q i,q i ) s.t. q i Q i and, at equilibrium, MC(q i,q i ) = 0 ( π = π( q))

18 Market: Equilibrium price: π Mixed Complementarity Model Agents problems min c i (q i,q i ) s.t. q i Q i and, at equilibrium, MC(q i,q i ) = 0 ( π = π( q)) Generalized Nash Game min c i (q i ) Agents problems s.t. q i Q i MC(q i, q i ) = 0 (same π i for all)

19 Market: Equilibrium price: π Mixed Complementarity Model Agents problems min c i (q i,q i ) s.t. q i Q i and, at equilibrium, MC(q i,q i ) = 0 ( π = π( q)) Generalized Nash Game Agents problems min c i (q i ) s.t. q i Q i MC(q i, q i ) = 0 (same π for all i) A Variational Equilibrium of the game is a Generalized Nash Equilibrium satisfying π i = π

20 Both models give same equilibrium Mixed Complementarity Model min c i (q i,q i ) Agents problems s.t. q i Q i and, at equilibrium, MC(q i,q i ) = 0 ( π = π( q)) Generalized Nash Game min c i (q i ) Agents problems s.t. q i Q i MC(q i, q i ) = 0 (same π for all i) J.P. Luna, C. Sagastizábal, M. Solodov. Complementarity and game-theoretical models for equilibria in energy markets: deterministic and risk-averse formulations. Ch. 10 in Risk Management in Energy Production and Trading, (R. Kovacevic, G. Pflüg and M. T. Vespucci), "Int. Series in Op. Research and Manag. Sci.", Springer, 2013.

21 Both models give same equilibrium Both models yield equivalent VIs Mixed Complementarity Model min c i (q i,q i ) Agents problems s.t. q i Q i and, at equilibrium, MC(q i,q i ) = 0 ( π = π( q)) Generalized Nash Game min c i (q i ) Agents problems s.t. q i Q i MC(q i, q i ) = 0 (same π for all i) J.P. Luna, C. Sagastizábal, M. Solodov. Complementarity and game-theoretical models for equilibria in energy markets: deterministic and risk-averse formulations. Ch. 10 in Risk Management in Energy Production and Trading, (R. Kovacevic, G. Pflüg and M. T. Vespucci), "Int. Series in Op. Research and Manag. Sci.", Springer, 2013.

22 Market: VI reformulation min c i (q i ) ith problem s.t. q i Q i MC(q i, q i ) = 0 Variational Inequality follows from optimality conditions

23 Market: VI reformulation ith problem min c i (q i ) s.t. q i Q i MC(q i, q i ) = 0 ( π) 1st order OC (primal form) q ic i ( q i ),q i q i 0 q i Q i MC Variational Inequality follows from optimality conditions

24 Market: VI reformulation ith problem min c i (q i ) s.t. q i Q i MC(q i, q i ) = 0 ( π i ) 1st order OC (primal form) q ic i ( q i ),q i q i 0 q i Q i MC Variational Inequality follows from optimality conditions In VI(F,C) : F( q),q q 0 feasible q the VI operator F(q) = the VI feasible set C = N F i (q) for F i (q) = q ic i (q i ) i=1 N Q i { } q : MC(q) = 0 i=1

25 Market: VI reformulation ith problem min c i (q i ) s.t. q i Q i MC(q i, q i ) = 0 ( π i ) 1st order OC (primal form) q ic i ( q i ),q i q i 0 q i Q i MC Variational Inequality follows from optimality conditions In VI(F,C) : F( q),q q 0 feasible q the VI operator F(q) = the VI feasible set C = N F i (q) for F i (q) = q ic i (q i ) i=1 N i=1 decomposability Q i { } q : MC(q) = 0

26 Market: VI reformulation ith problem min c i (q i ) s.t. q i Q i MC(q i, q i ) = 0 ( π i ) 1st order OC (primal form) q ic i ( q i ),q i q i 0 q i Q i MC Variational Inequality follows from optimality conditions In VI(F,C) : F( q),q q 0 feasible q the VI operator F(q) = the VI feasible set C = N F i (q) for F i (q) = q ic i (q i ) i=1 N i=1 decomposability Q i { } q : MC(q) = 0 NOTE: MC does not depend on i: constraint is shared

27 Incorporating a Capacity Market Suppose producers pay I i (z i ) to invest in an increase z i in production capacity Production bounds go from 0 q i q i max ( q i Q i ) to 0 q i q i max+z i (z i,q i ) X i

28 Incorporating a Capacity Market Suppose producers pay I i (z i ) to invest in an increase z i in production capacity Production bounds go from 0 q i q i max ( q i Q i ) to 0 q i q i max + z i (z i,q i ) X i

29 Incorporating a Capacity Market Suppose producers pay I i (z i ) to invest in an increase z i in production capacity Production bounds go from 0 q i q i max ( q i Q i ) to 0 q i q i max + z i (z i,q i ) X i min I i (z i ) + c i (q i ) mini i (z i ) + V i (z i ) { ith problem s.t. (z i,q i ) X i V (z i minc i (q i ) ) = MC(q i,q i (z i,q i ) X i ) = 0 MC(q i,q i ) = 0

30 Incorporating a Capacity Market Suppose producers pay I i (z i ) to invest in an increase z i in production capacity Production bounds go from 0 q i q i max ( q i Q i ) to 0 q i q i max + z i (z i,q i ) X i min I i (z i ) + c i (q i ) mini i (z i ) + V i (z i ) { ith problem s.t. (z i,q i ) X i V (z i minc i (q i ) ) = MC(q i,q i (z i,q i ) X i ) = 0 MC(q i,q i ) = 0

31 Incorporating a Capacity Market Suppose producers pay I i (z i ) to invest in an increase z i in production capacity Production bounds go from 0 q i q i max ( q i Q i ) to 0 q i q i max + z i (z i,q i ) X i min I i (z i ) + c i (q i ) mini i (z i ) + V i (z i ) { ith problem s.t. (z i,q i ) X i V (z i minc i (q i ) ) = MC(q i,q i (z i,q i ) X i ) = 0 can this problem be rewritten as a 2-level problem? MC(q i,q i ) = 0

32 Incorporating a Capacity Market When trying to rewrite min I i (z i ) + V i (z i ) using minc i (q i ) V i (z i ),q i ) = (z i,q i ) X i MC(q i,q i ) = 0 a difficulty arises.

33 Incorporating a Capacity Market When trying to rewrite min V i (z i, q i ) = I i (z i ) + V i (z i ) using minc i (q i ) (z i,q i ) X i MC(q i, q i ) = 0 a difficulty arises. The function V i depends on (z i,q i ), the second-level problem is a Generalized Nash Game (hard!)

34 Incorporating a Capacity Market When trying to rewrite min V i (z i,q i ) = I i (z i ) + V i (z i ) using minc i (q i ) (z i,q i ) X i MC(q i,q i ) = 0 a difficulty arises. The function V i depends on (z i,q i ), the second-level problem is a Generalized Nash Game (hard!) Consistent with reality: Agents will keep competing after capacity expansion

35 Incorporating a Capacity Market When trying to rewrite min V i (z i,q i ) = I i (z i ) + V i (z i ) using minc i (q i ) (z i,q i ) X i MC(q i,q i ) = 0 a difficulty arises. The function V i depends on (z i,q i ), the second-level problem is a Generalized Nash Game (hard!) Consistent with reality: Agents will keep competing after capacity expansion. Similarly for Mixed Complementarity model

36 Incorporating a Capacity Market When trying to rewrite min V i (z i,q i ) = I i (z i ) + V i (z i ) using minc i (q i ) (z i,q i ) X i MC(q i,q i ) = 0 a difficulty arises. The function V i depends on (z i,q i ), the second-level problem is a Generalized Nash Game (hard!) Consistent with reality: Agents will keep competing after capacity expansion. Similarly for Mixed Complementarity model and 2 stage with recourse, even without expansion

37 What about uncertainty? Given k = 1,..., K uncertain scenarios (demand, costs, etc) Investment variables are (naturally) the same for all realizations: z i Production variables are (naturally) different for each realization: q i k ith problem for scenario k min I i (z i ) + c i k (q k) s.t. (z i,q i k ) Xi k MC k (q i k,q i k ) = 0 Two-stage formulation with recourse not possible

38 What about uncertainty? Given k = 1,..., K uncertain scenarios (demand, costs, etc) Investment variables are (naturally) the same for all realizations: z i Production variables are (naturally) different for each realization: q i k ith problem for scenario k min I i (z i ) + c i k (q k) s.t. (z i,q i k ) Xi k MC k (q i k,q i k ) = 0 Two-stage formulation with recourse not possible Single-stage formulation instead: find a capacity expansion compatible with K scenarios of competition.

39 What about uncertainty? Given k = 1,..., K uncertain scenarios (demand, costs, etc) Investment variables are (naturally) the same for all realizations: z i Production variables are (naturally) different for each realization: q i k ith problem for scenario k min I i (z i ) + c i k (q k) s.t. (z i,q i k ) Xi k MC k (q i k,q i k ) = 0 Two-stage formulation with recourse not possible Single-stage formulation instead: find a capacity expansion compatible with K scenarios of competition (likewise for generation-only market)

40 Which Stochastic VI? Risk-neutral agents Derive VI from ith problem using expected value min I i (z i ) + E[c i k (qi k )] s.t. (z i,q i k ) Xi k for k = 1 : K MC k (q i k,q i k ) = 0 for k = 1 : K

41 Which Stochastic VI? Risk-neutral agents Derive VI from ith problem min I i (z i ) + E[c i k (qi k )] s.t. (z i,q i k using expected value ) Xi k for k = 1 : K MC k (q i k,q i k ) = 0 for k = 1 : K ] a VI operator F involving I i (z i ) q i E [c i 1:K 1:K (q) a VI feasible set C = K k=1 N i=1 X i k { qk : MC k (q k ) = 0 }

42 Which Stochastic VI? Risk-neutral agents Derive VI from ith problem min I i (z i ) + E[c i k (qi k )] s.t. (z i,q i k using expected value ) Xi k for k = 1 : K MC k (q i k,q i k ) = 0 for k = 1 : K ] a VI operator F involving I i (z i ) q i E [c i 1:K 1:K (q) a VI feasible set C = K k=1 N i=1 X i k { qk : MC k (q k ) = 0 } decomposability

43 Which Stochastic VI? Risk-neutral agents Derive VI from ith problem min I i (z i ) + E[c i k (qi k )] s.t. (z i,q i k using expected value ) Xi k for k = 1 : K MC k (q i k,q i k ) = 0 for k = 1 : K ] a VI operator F involving I i (z i ) q i E [c i 1:K 1:K (q) K N a VI feasible set C = X i k { qk : MC k (q k ) = 0 } k=1 i=1 decomposability there is no coupling between scenarios (E is linear)

44 Which Stochastic VI? Risk-averse agents, risk measure ρ Derive VI from ith problem using risk measure min I i (z i ) + ρ[c i k (qi k )] s.t. (z i,q i k ) Xi k for k = 1 : K MC k (q i k,q i k ) = 0 for k = 1 : K

45 Which Stochastic VI? Risk-averse agents, risk measure ρ Derive VI from ith problem using risk measure min I i (z i ) + ρ[c i k (qi k )] s.t. (z i,q i k ) Xi k for k = 1 : K MC k (q i k,q i k ) = 0 for k = 1 : K Difficulties arise: The risk measure is in general nonsmooth ρ(z) := AV@R ε (Z) = min u {u ε E ([Z k u] +)} : it is a value-function and [ ] + is nonsmooth

46 Which Stochastic VI? Risk-averse agents, risk measure ρ Derive VI from ith problem using risk measure min I i (z i ) + ρ[c i k (qi k )] s.t. (z i,q i k ) Xi k for k = 1 : K MC k (q i k,q i k ) = 0 for k = 1 : K Difficulties arise: The risk measure is in general nonsmooth ρ(z) := AV@R ε (Z) = min u {u + 1 ε ([Z 1 E k u] +)} : it is a value-function and [ ] + is nonsmooth ] the VI operator F involves I i (z i ) q i ρ [c i 1:K 1:K (q), multivalued

47 Two ways of handling multivalued VI operator Reformulation: Introduce directly into the agent problem, by rewriting [] + in ρ(z) := min u { u ε E ( [Z k u] +)} by means of new variables and constraints

48 Two ways of handling multivalued VI operator Reformulation: Introduce directly into the agent problem, by rewriting [] + in ρ(z) := min u { u ε E ( [Z k u] +)} by means of new variables and constraints Smoothing: Smooth the [ ] + -function and solve the smoothed VI { ρ l (Z) := min u + 1 ( u 1 ε E σ l (Z k u)) }, for smoothing σ l [ ] + uniformly as l

49 Reformulation ( ρ(z) = min u {u + 1 ε 1 [Z k u] +)} min I i (z i ) + ρ[c i k (qi k )] FROM s.t. (z i,q i k ) Xi k for k = 1 : K TO: MC k (q i k,q i k ) = 0 for k = 1 : K min I i (z i ) + u i ε E (T i k s.t. ) (z i,q i k ) Xi k for k = 1 : K MC k (q i k,q i k ) = 0 for k = 1 : K Tk i ci k (qi k ) ui,tk i 0 for k = 1 : K,u IR

50 Reformulation ( ρ(z) = min u {u + 1 ε 1 [Z k u] +)} min I i (z i ) + ρ[c i k (qi k )] FROM s.t. (z i,q i k ) Xi k for k = 1 : K TO: MC k (q i k,q i k ) = 0 for k = 1 : K min I i (z i ) + u i ε E (T i k s.t. ) (z i,q i k ) Xi k for k = 1 : K MC k (q i k,q i k ) = 0 for k = 1 : K Tk i ci k (qi k ) ui,tk i 0 for k = 1 : K,u IR NOTE: new constraint is NOT shared: no longer a generalized Nash game, but a bilinear CP (how to show?).

51 Assessing both options PATH can be used for the two variants. + Reformulation eliminates nonsmoothness Non-separable feasible set + Smoothing To drive smoothing parameter to 0: repeated VI solves Keeps feasible set separable by scenarios: easier VI

52 Assessing both options PATH can be used for the two variants. + Reformulation eliminates nonsmoothness Non-separable feasible set + Smoothing To drive smoothing parameter to 0: repeated VI solves Keeps feasible set separable by scenarios: easier VI Provides existence result!

53 Smoothing We use smooth approximations ρ l ρ l (Z) := min u { u ε E[ σ l (Z k u) ] } for smoothing σ l [ ] + uniformly as l. For instance, σ l (t) = (t + t 2 + 4τ 2 l )/2 for τ l 0. Since ρ l is smooth, VI(F l [,C) has a a single-valued ] VI operator involving q iρ l (c i k (q k)) K k=1,

54 Theorems like ρ l is a risk-measure convex, monotone, and translation equi-variant, but not positively homogeneous (only coherent in the limit). ρ l is C 2 for strictly convex smoothings such as σ l (t) = (t + t 2 + 4τ 2 l )/2 Any accumulation point of the smoothed problems solves the original risk-averse (non-smooth) problem as l. : existence result!

55 Theorems like ρ l is a risk-measure convex, monotone, and translation equi-variant, but not positively homogeneous (only coherent in the limit). ρ l is C 2 for strictly convex smoothings such as σ l (t) = (t + t 2 + 4τ 2 l )/2 Any accumulation point of the smoothed problems solves the original risk-averse (non-smooth) problem as l. existence result! Reference: An approximation scheme for a class of risk-averse stochastic equilibrium problems. Luna, Sagastizábal, Solodov

56 Numerical performance of smoothing τ l VI l τ l+1 VI l+1...until stabilization x for x = (z 1:N,q 1:N 1:K ) stop if j+1 x j max (1, ) 0.01 x j+1 2 players and a consumer representative, player 0. Player 2 has higher generation costs. Less than 5 solves in average, each solve takes 45 seconds. Excellent solution quality

57 Numerical performance of smoothing τ l VI l τ l+1 VI l+1...until stabilization x for x = (z 1:N,q 1:N 1:K ) stop if j+1 x j max (1, ) 0.01 x j+1 2 players and a consumer representative, player 0. Player 2 has higher generation costs. Less than 5 solves in average, each solve takes 45 seconds. Excellent solution quality

58 Numerical performance of smoothing τ l VI l τ l+1 VI l+1...until stabilization x for x = (z 1:N,q 1:N 1:K ) stop if j+1 x j max (1, ) 0.01 x j+1 2 players and a consumer representative, player 0. Player 2 has higher generation costs. Less than 5 solves in average, each solve takes 45 seconds. Excellent solution quality After the first VI, PATH solves much faster VIs for l > 1

59 Numerical performance of smoothing For nonconvex generation costs, reformulation becomes slower with nonconvex generation costs. Smoothing needs less than 6 solves in average. Once again, after the first VI solve, PATH much faster for consecutive smoothed VIs: time of PATH smoothing 2 time of PATH reformulation but: Total time of reformulation increases a lot, it scales less well

60 Numerical performance of smoothing For nonconvex generation costs, reformulation becomes slower with nonconvex generation costs. Smoothing needs less than 6 solves in average. Once again, after the first VI solve, PATH much faster for consecutive smoothed VIs: time of PATH smoothing 2 time of PATH reformulation but: Total time of reformulation increases a lot, it scales less well

61 Final Comments When in the agents problems the objective or some constraint depends on actions of other agents, writing down the stochastic game/vi can be tricky (which selection mechanism in a 2-stage setting?) Handling nonsmoothness via reformulation seems inadequate for large instances Smoothing solves satisfactorily the original risk-averse nonsmooth problem for moderate τ (no need to make τ 0) Smoothing preserves separability; it is possible to combine Benders techniques (along scenarios) with Dantzig-Wolfe decomposition (along agents) Decomposition matters: for European Natural Gas network Solving VI directly with PATH solver S. Dirkse, M. C. Ferris, and T. Munson Using DW-decomposition saves 2/3 of solution time

62 SAVE THE DATES! June 25th-July 1st, 2016